|
--- |
|
language: |
|
- en |
|
tags: |
|
- pytorch |
|
- text-generation |
|
- causal-lm |
|
- rwkv |
|
license: apache-2.0 |
|
datasets: |
|
- the_pile |
|
|
|
--- |
|
|
|
# RWKV-4 1.5B |
|
|
|
# WARNING: 4 / 4a / 4b models ARE NOT compatible. Use RWKV-4 models unless you know what you are doing. |
|
# WARNING: 4 / 4a / 4b models ARE NOT compatible. Use RWKV-4 models unless you know what you are doing. |
|
# WARNING: 4 / 4a / 4b models ARE NOT compatible. Use RWKV-4 models unless you know what you are doing. |
|
|
|
## Model Description |
|
|
|
RWKV-4 1.5B is a L24-D2048 causal language model trained on the Pile. See https://github.com/BlinkDL/RWKV-LM for details. |
|
|
|
Use https://github.com/BlinkDL/ChatRWKV to run it. |
|
|
|
ctx_len = 1024 |
|
n_layer = 24 |
|
n_embd = 2048 |
|
|
|
RWKV-4-Pile-1B5-20220929-ctx4096.pth : Fine-tuned to ctx_len 4096. |
|
* Likely better when ctxlen > 100. Please test. |
|
|
|
RWKV-4-Pile-1B5-20220903-8040.pth : Trained on the Pile for 332B tokens. |
|
* Pile loss 2.0415 |
|
* LAMBADA ppl 7.04, acc 56.43% |
|
* PIQA acc 72.36% |
|
* SC2016 acc 68.73% |
|
* Hellaswag acc_norm 52.48% |
|
|
|
### Instruct-test models: only useful if you construct your prompt following dataset templates |
|
|
|
Note I am using "Q: instruct\n\nA: result" prompt for all instructs. |
|
|
|
RWKV-4-Pile-1B5-Instruct-test1-20230124.pth |
|
instruct-tuned on https://huggingface.co/datasets/bigscience/xP3all/viewer/en/train |
|
|
|
RWKV-4-Pile-1B5-Instruct-test2-20230209.pth |
|
instruct-tuned on https://huggingface.co/datasets/Muennighoff/flan & NIv2 |
|
|
|
### Chinese models |
|
|
|
RWKV-4-Pile-1B5-EngChn-testNovel-xxx for writing Chinese novels (trained on 200G Chinese novels.) |
|
|
|
RWKV-4-Pile-1B5-EngChn-testxxx for Chinese Q&A (trained on 10G Chinese text. only for testing purposes.) |
|
|
|
## Note: 4 / 4a / 4b models ARE NOT compatible. Use RWKV-4 unless you know what you are doing. |
|
|
|
RWKV-4b-Pile-1B5-20230217-7954.pth (--my_testing 'a') with tiny amt of QKV attention to improve performance |
|
* Pile loss 1.9947 |
|
* LAMBADA ppl 5.82, acc 62.35% |
|
* PIQA acc 72.52% |
|
* SC2016 acc 68.89% |
|
* Hellaswag acc_norm 54.32% |
|
|
